Learning to Trigger: Reinforcement Learning at the Large Hadron Collider

High-throughput scientific facilities like the Large Hadron Collider (LHC) rely on real-time event filtering, known as triggering, to manage massive data streams. Traditionally, these trigger menus are static and hand-tuned, which makes them vulnerable to performance degradation as detector conditions, pileup, and background compositions drift over time. This paper investigates whether a reinforcement learning (RL) agent can replace these static systems by dynamically adjusting trigger thresholds in real time to maintain stable background rates while maximizing signal efficiency.

The researchers frame online threshold tuning as a sequential decision-making problem. They introduce a framework that uses a recurrent neural network to ingest streaming summaries of recent event rates and physics-informed features. To address the strict operational constraints of the LHC, they adapt Group-Filtered Policy Optimization (GFPO) to the streaming setting, introducing two variants—GFPO-F and GFPO-FR—that enforce background rate feasibility during training by filtering candidate actions based on their adherence to target bandwidths.

The RL-based controller significantly outperforms traditional PID and non-learning baselines. On Monte Carlo simulations, the agent increased the fraction of time intervals where the background rate remained within the target tolerance by up to 48% for energy-based triggers and 28% for anomaly-detection triggers. Crucially, when deployed on real CMS collision data without any fine-tuning, the agent maintained this superior performance, achieving a 56% improvement in rate stability for energy-based triggers and a 28% improvement for anomaly-detection triggers, while simultaneously increasing signal efficiency.

This work represents the first successful demonstration of an RL-based trigger control system on real LHC collision data. By moving away from static, hand-tuned menus toward autonomous, adaptive systems, this approach offers a scalable solution for managing the increasing data complexity of modern particle physics experiments. The success of the GFPO variants in this domain also suggests that these techniques for constrained, streaming decision-making may generalize to other scientific and industrial applications where real-time control must balance performance against strict safety or resource constraints.